MULTI-AGENT COALITION FORMATION

BASED ON CLONAL SELECTION

Martina Husáková

Department of Information Technologies, University of Hradec Králové, Hradecká 1249, Hradec Králové, Czech Republic

Keywords: Biological Immune System, Artificial Immune System, Multi-Agent System, Coalition Formation.

Abstract: Coalition formation is by one of cooperative mechanisms used in multi-agent systems. Various approaches

of multi-agent coalition formation ensure the coalition stability, optimal allocation of sources and payoff

distribution for efficient achievement of individual or collective goals. The paper investigates mechanisms

of human immune system from the coordination and cooperation point of view. It specifies requirements for

effective multi-agent coalition formation on the basis of these mechanisms. The clonal selection-based

algorithm is used to discover the optimal coalition structure. Finally, further directions in the research of

exploitation artificial immune-based algorithms for multi-agent coalition formation are mentioned.

1 INTRODUCTION

Specific activities require a certain deal of

coordination or cooperation of more artificial

autonomous agents to achieve goals with minimal

spend time and costs. The attention is paid to the

efficient coordination and cooperation of artificial

autonomous agents in multi-agent systems (MAS).

The multi-agent coalition formation is one of

cooperative mechanisms intensively investigated in

the MAS. The paper evaluates the human biological

immune system (BIS) from the coordination and

cooperation points of view and specifies

requirements of the effective multi-agent coalition

formation based on this approach. The paper

proposes the use of the immune-based algorithm

ClonAlg for generating the optimal coalition

structure.

2 MULTI-AGENT COALITION

FORMATION

The coalition formation is a type of cooperation of

artificial autonomous agents in the MAS. The

coalition is a goal-oriented and short-lived group of

agents solving a specific pre-defined problem. Any

cooperation with artificial autonomous agents from

others relevant information by software web agents

is another example. The agent knowledge base can

be extended by communication with others agents

which provide new information to other agents. If

they join together, they can cover wider search space

of web documents.

Finding the optimal coalition structure is one of

the problems solved in the multi-agent coalition

formation. The coalition structure is a group of

various coalition combinations that are able to solve

a specific problem. The question is how to find such

a coalition structure containing agents able to use

their sources optimally. Number of coalition

structures increases exponentially with the number

of agents. The goal is to find efficient algorithms

that are able to search solutions with minimal

computational sources and time requirements.

Dynamic programming, genetic algorithms or

greedy algorithms are usually used for this purpose

(Rahwan et al., 2009).

The problem of optimal coalition structure

generation has not been sufficiently researched by

the different biology-inspired approach resulting

from properties and mechanisms of the human

biological immune system (BIS). The BIS can offer

new methods for solving problems in multi-agent

coalition formation. The algorithm of the clonal

selection inspired by the clonal selection principle is

investigated for the finding of the optimal coalition

structure.

231

Husáková M..

MULTI-AGENT COALITION FORMATION BASED ON CLONAL SELECTION.

DOI: 10.5220/0003572202310234

In Proceedings of the 8th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2011), pages 231-234

ISBN: 978-989-8425-74-4

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

3 THE IMMUNE SYSTEM AND

COALITION FORMATION

The BIS is a complex system that maintains the

homeostasis in the living organism. It is able to

recognize dangerous objects (antigens) that invoke

the immune response. Antigen can be a part of the

organism and risky for it (self-antigen) or can come

from the outer environment, e. g. viruses, bacteria or

fungi (non-self-antigen).

The BIS includes the innate and adaptive

immunity. The innate immunity guards the

organism. It is supported by mechanic barriers and

different types of immune cells. The adaptive

immunity focuses on recognition of already

identified antigens. B-lymphocytes (B-ly) and T-

lymphocytes (T-ly) are the main representatives of

this layer. The B-ly is the main producer of

antibodies and plays the role of memory cell. The T-

ly eliminates dangerous antigens or regulates

functions of others immune cells (Castro and

Timmis, 2002).

There exist similarities between formation of

immune cells into groups and the group behaviour

necessary for solving complex problems of artificial

autonomous agents. The immune cells are able to

join and cooperatively solve a specific problem on

the basis of specific stimuli. They use cellular

signalling pathways that ensure transferring pieces

of information from the source to the receiver. The

immune cells can be perceived as pro-social

biological agents that keep track of interests of the

whole the BIS. Generation of antibodies, activities

of complement system, phagocytosis or cytokines

production are demonstrative examples of

cooperation activities in case of the BIS.

The following table mentions key properties for

the stability maintenance of the BIS (Dasgupta and

Nino, 2008). The same attributes are specified as

requirements for effective multi-agent coalition

formation in this paper.

The conjunctions and similarities, displayed in

the table, show the relevance for deeper research of

the mutually utility of these two systems. The key

question is how the concrete mechanisms of BIS can

be used for discovering the optimal coalition

structure. Artificial immune systems research area

offers collection of immune-based algorithms that

should be researched for this purpose.

Table 1: Coalition formation in the view of the BIS.

BIS attributes Coalition requirements

Distributiveness

This attribute occurs e. g. in

case of lymphocytes

production by the bone

marrow in different places of

the organism. Positive

selection, negative selection,

clonal selection and immune

network concept are

characterized by the

distributiveness, too.

Distributiveness

If the agent fails in filling the

task, despite this the task

should be completed. The

coalition formation should not

be directed by one central

agent of the MAS.

Communication

The BIS cells communicate

directly or indirectly with

the environment. Direct

communication is executed

mainly with the aid of

adhesive molecules.

Cytokines are products of

immune cells activities that

ensure the indirect

communication in the BIS.

Communication

Communication is direct (e. g.

protocols) or indirect

(stigmergy). It helps in

receiving information from the

environment, e. g. agents or

others coalitions.

Robustness and stability

Robustness and stability are

expressed in cooperation of

different immune cells and

organs. These processes lead

to the emergent property of

the BIS.

Robustness and stability

Agents should monitor their

states and atmosphere in the

coalition. If something is

wrong, the agent signalizes it

to the others. Agents should

be motivated to stay in the

coalition during the mission

fulfilment.

Dynamics and adaptation

Dynamics and adaptation are

expressed in cooperation of

different immune cells and

organs. These processes lead

to the emergent property of

BIS.

Dynamics and adaptation

The coalition of agents should

behave flexibly in case of

changing environment.

Learning and memory

Immune cells are able to

learn with the aid of

feedback received from the

environment and to

remember already identified

antigens. This ensures faster

reaction of immune cells in

the future. Memory is

formed with the aid of

immune network, gene

libraries or memory cells.

Learning and memory Agent

should be able to learn and use

the learned in the future.

Agent´s memory helps to react

adequately to stimuli.

4 ARTIFICIAL IMMUNE

SYSTEMS

Artificial immune systems (AIS) are the research

subarea of computational intelligence that is inspired

by the BIS behavior. In present, four groups of

artificial immune system algorithms are used.

The first one stems from the diversity generation

of immune cells with the aid of gene libraries

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

232

(Castro and Timmis, 2002). These libraries contain

gene fragments combined for creation different

immune cells. There are artificial gene libraries in

AIS. These gene libraries are used for generating

potential solutions changed by the mutation process

towards the approximation to the optimal solution.

The second one is the group of selected

algorithms – the algorithm of positive selection,

algorithm of negative selection, algorithm of clonal

selection and their modifications (Dasgupta and

Nino, 2008), (Castro, 2006), (Castro and Timmis,

2002). The algorithms of positive and negative

selection are mainly applied to classification and

recognition problems. The computer security or fault

detection is a typical application. The algorithm of

clonal selection is used especially for classification

and optimization problems (Castro and Zuben,

2000).

The third group is inspired by the immune

network theory by N. K. Jerne (Jerne, 1974). This

theory apprehends the BIS as a network of

interconnected stimulated B-cells interacting with

each other. Continuous artificial immune networks

are used mainly for modeling and simulation of the

BIS with the aid of differential or difference

equations. Discrete artificial immune networks are

based on differential equations or iterative

procedures. They are used mainly for the pattern

recognition, data analysis, machine learning or

optimization problems.

The dendritic cell algorithm is part of the last

group. It has already been used for anomalies

detection in a computer network as a classifier for

scanning computer ports (Greensmith, et al., 2005).

5 ClonAlg FOR COALITION

FORMATION

ClonAlg is the clonal selection-based algorithm

inspired by the clonal selection principle explaining

the process of antibody generation. If a B-ly

recognizes the antigen, clones of the same

specificity are created. Mutation occurs during the

cloning. It can improve the affinity (tightness of

bond) between the antigen and the B-ly in the future

reunion. The ClonAlg was originally designed for

the pattern recognition. The optimized version of

ClonAlg (ClonAlg-opt) is used for optimization

tasks (de Castro and von Zuben, 2002). The

ClonAlg-opt uses population of antibodies. This

population pictures potential solutions of the

problem (set P). The pseudo-code of the algorithm

follows this procedure (de Castro, 2006):

1. Initialization: create an initial population of

antibodies (P).

2. Fitness evaluation: determine the fitness of

each element of P.

3. Clonal selection and expansion: select n

highest fitness elements of P and generate

clones of these antibodies proportionally to

their fitness: the higher the fitness, the higher

the number of copies, and vice-versa.

4. Affinity maturation: mutate all these copies

with a rate that is inversely proportional to

their fitness: the higher the fitness, the

smaller the mutation rate, and vice-versa.

Add these mutated individuals to the

population P.

5. Meta-dynamics: replace a number n

of low

fitness individuals by (randomly generated)

new ones.

6. Cycles: repeat step 2 to 5 until a certain

stopping criterion is met.

The usage of the ClonAlg-opt for generating the

optimal coalition structure is designed for the

problem of elimination of oil spills. This problem

requires the efficient, fast and optimal cooperation

of sources because oil spills are spread in dynamic

environment.

Three agents eliminate oil spills of two different

oils in this demonstrative example. Every agent has

the list of two properties relevant for elimination of

oil spills. The first one describes how the agent

eliminates the portion of the first type of oil spill, i.

e. the usefulness of agent’s sensor in the detection of

the first type of oil spill. The second one describes

how the agent eliminates the portion of the second

type of oil spill. The coalition structure consists of

different coalitions and is represented by the

antibody molecule. The coalition consists of one or

more agents identified by an identification value.

Permutation encoding is used for the coalition

representation in the coalition structure. Two

restrictions have to be respected: Every oil spill has

to be refined by one agent minimally and coalitions

have not overlap themselves. The pseudo-code of

the algorithm follows:

1. Initialization: create an initial population of N

coalition structures (set P), eliminate

overlapping and empty coalitions.

2. Fitness evaluation: determine the quality

(fitness) of N coalition structures.

3. Clonal selection and expansion: select n

highest fitness coalition structures (e. g.

MULTI-AGENT COALITION FORMATION BASED ON CLONAL SELECTION

233

fitness-proportionate selection, stochastic

universal sampling or tournament selection

can be used) and generate clones of these

ones proportionally to their fitness.

4. Affinity maturation: mutate all these copies

with a rate that is inversely proportional to

their fitness. Add these mutated coalition

structures to the population P. The inversion

operator is suitable for it. We change the

position of agents of coalition and influence

the quality of coalition that eliminates one

particular oil spill.

5. Meta-dynamics: replace a number n

of low

fitness coalition structures by (randomly

generated) new ones.

6. Cycles: repeat step 2 to 5 until a certain final

criterion is met.

6 CONCLUSIONS

The paper deals with the coalition formation of

artificial autonomous agents of the MAS in the

context of the BIS behavior. Key properties and

mechanisms of the BIS cooperation are identified

and used for the requirements specification for

multi-agent coalitions.

The paper proposes to use the ClonAlg-opt

algorithm for generating the optimal coalition

structure. Experiments are going to be realized with

the ClonAlg-opt algorithm and other clonal

selection-based AIS algorithms. The immune

network-based algorithms Opt-aiNet and Dopt-aiNet

(Dasgupta and Nino, 2008) are going to be used for

the coalition formation, too. They stem from the data

clustering aiNet algorithm (Castro and Zuben, 2001)

and are mainly used for the optimization purposes.

Results of these experiments will be compared to the

genetic algorithms. The multi-agent modeling or

simulation tool will be used for experiments.

NetLogo, Webots or Anylogic are suitable

candidates for the above mentioned purposes.

ACKNOWLEDGEMENTS

This paper was supported by the project N.

P403/10/1310 “SMEW - Smart Environment at

Workplaces” of Czech Science Foundation and FIM

UHK Internal Grant “Immunity-based Multi-Agent

Coalition Formation” N. 12/2011.

REFERENCES

Dasgupta, D., Nino, F., 2008. Immunological

Computation: Theory and Applications. Auerbach

Publication. 1

edition. p. 296. ISBN 978-

1420065459.

de Castro, L. N., 2006. Fundamentals of natural

computing: basic concepts, algorithms, and

applications. Chapman and Hall/CRC, 1

Ed, p. 696.

ISBN 978-1584886433.

de Castro, L. N., von Zuben, F. J., 2002. The Clonal

Selection Algorithm With Engineering Application. In

Proceedings of the Genetic and Evolutionary

Computation Conference (GECCO). Morgan

Kaufmann. pp. 36 – 37.

de Castro, L. N., Timmis, J., 2002. Artificial Immune

Systems: A New Computational Intelligence

Approach. Springer-Verlag, 1

edition. p. 357. ISBN

978-1852335946.

de Castro, L., von Zuben, F. J., 2001. Data Mining: A

Heuristic Approach. Idea Group Publishing, Chapter

XII, aiNet: An Artificial Immune Network for Data

Analysis. pp. 231 – 259.

Greensmith, J., et al., 2005. Introducing Dendritic Cells as

a Novel Immune-inspired Algorithm for Anomaly

Detection. In Proceedings of the 4

International

Conference on Artificial immune systems (ICARIS).

Springer-Verlag. LNCS, Vol. 3627, pp. 153-167.

Horling, B., Lesser, V., 2005. A Survey of Multi-Agent

Organizational Paradigms. The Knowledge

Engineering Review. 19(4): pp. 281-316. The

Cambridge University Press.

Jerne, N. K., 1974. Towards a Network Theory of the

Immune System. In An Immunologica. 125C(1-2): pp.

373-389.

Rahwan, T., et al., 2009. An anytime algorithm for

optimal coalition structure generation. In Journal of

Artificial Intelligence Research (JAIR). 34(1): pp.

521-567.

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

234